BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:65@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20240725T100000 DTEND;TZID=Asia/Kolkata:20240725T110000 DTSTAMP:20240716T030922Z URL:https://cds.iisc.ac.in/events/ph-d-thesis-defense-cds-25-july-2024-fas t-and-scalable-algorithms-for-intelligent-routing-of-autonomous-marine-veh icles/ SUMMARY:Ph.D. Thesis Defense: CDS: 25\, July 2024 "Fast and Scalable Algori thms for Intelligent Routing of Autonomous Marine Vehicles" DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Def ense\n\n\n\nSpeaker : Mr. Rohit Chowdhury\nS.R. Number : 06-18-01-10-12-18 -1-16320\nTitle : "Fast and Scalable Algorithms for Intelligent Routing of Autonomous Marine Vehicles"\nThesis examiner: Dr. Indranil Saha\, IIT Kan pur\nResearch Supervisor: Dr. Deepak Subramani\nDate &\; Time : July 25 \, 2024 (Thursday) at 10:00 AM\nVenue : # 102 CDS Seminar Hall\n\n\n\nABST RACT\n\nAutonomous marine agents play a pivotal role in diverse ocean appl ications. These agents serve as indispensable instruments for acquiring cr ucial environmental information. They are used to explore and monitor of h arsh environments\, e.g.\, to map ocean topography\, study coral reefs\, s earch and rescue\, structural monitoring of oil and gas installations etc. In naval security\, these agents are used for surveillance and strategic monitoring of maritime activities. Building intelligence to optimally use these agents is essential for reducing operational costs.\n\nThe path plan ning problem is as follows. An autonomous marine agent must optimally trav erse from a given start location to a given target location in a stochasti c dynamic velocity field like ocean currents while avoiding obstacles or r estricted regions in the flow. A key challenge is that the agent is heavil y advected by the flow. The optimality may refer to minimising expected tr avel time or energy consumption\, data collection or risk of failure. Whil e there are multiple methods of solving path planning problems\, each with its challenges\, we develop and use a fast and scalable MDP-based offline planning software that computes optimal policies\, and a novel sequence-m odelling-based deep learning approach for onboard routing and dynamic plan ning\, where the objective is to learn optimal action sequences for the ag ent. The goal of this thesis is to develop efficient\, fast and scalable A rtificial intelligence algorithms for optimal planning and on-board routin g algorithms for autonomous marine agents in stochastic dynamic environmen ts.\n\nThe thesis comprises five works organised into two parts based on t he solution approach. In the first part\, we model the path planning probl em as a Markov Decision Process (MDP) and aim to compute an optimal policy . However\, the key challenge here is that solving an MDP can be prohibiti vely expensive for large state and action spaces. To overcome this challen ge\, we either approximate the optimal policy or accelerate the computatio n using GPUs.\n\n Physics-driven Q-learning for onboard routing: First\, the distribution of exact time-optimal paths predicted by stochastic Dynam ically Orthogonal (DO) Hamilton-Jacobi level set partial differential equa tions (HJLS PDEs) are utilised to learn an initial action-value function t hat approximates the optimal policy. The flow data collected by onboard se nsors are utilised to get a posterior estimate of the environment. The app roximated optimal policy is refined in-mission by performing epsilon greed y Q-learning in simulated posterior environments. We showcase the computat ional advantage of the approach at the cost of approximating the optimal p olicy.\n GPU-accelerated path planning: We compute an exact optimal polic y by solving the path planning problem modelled as an MDP. To solve large- scale real-time problems\, which can otherwise be computationally expensiv e\, we introduce an efficient end-to-end GPU accelerated algorithm that bu ilds the MDP model (computing transition probabilities and expected one-st ep rewards) and solves the MDP to compute an optimal policy. We develop me thodical and algorithmic solutions to overcome the limited global memory o f GPUs by using a dynamic reduced-order representation of the ocean flows\ , leveraging the sparse nature of the state transition probability matrix and introducing a neighbouring subgrid concept to save memory and reduce t he computational effort. We achieve significant speedups compared to conve ntional sequential computation.\n Multi-objective GPU-accelerated path pl anning: The end-to-end GPU accelerated MDP solver is extended to a multi-o bjective path planner to solve multi-objective optimisation problems in pa th planning\, like minimising both the expected mission completion time an d energy consumption. MDPs are modelled with scalarised rewards for multip le objectives. The solver is used to solve numerous instances of complex s cenarios with other sources of uncertainty in the environment\, enabling u s to compute optimal operating curves in a fraction of the time compared t o traditional solvers.\n\nIn the second part\, we convert the optimal path planning problem into a supervised learning problem via sequence modellin g. This approach involves learning optimal action sequences based on the a vailable environment information and expert trajectories. It also avoids t he issue of re-computing optimal policies for onboard routing.\n Intellig ent onboard routing using decision transformers: We develop a novel\, deep learning method based on the decision transformer (decoder-only model) fo r onboard routing of autonomous marine agents. Training data is obtained f rom aforementioned HJLS PDE or MDP solvers\, which is further processed to sequences of states\, actions and returns. The model is autoregressively trained on these sequences and then tested in different environment settin gs. We demonstrate that (i) a trained agent learns to infer the surroundin g flow and perform optimal onboard routing when the agent's state estimati on is accurate\,(ii) specifying the target locations (in case of multiple targets) as a part of the state enables a trained agent to route itself to the correct destination\, and (iii) a trained agent is robust to limited noise in state transitions and is capable of reaching target locations in completely new flow scenarios. We extensively showcase end-to-end planning and onboard routing in various canonical and idealised ocean flow scenari os.\n Path planning with environment encoders and action decoders: We pro pose a novel combination of dynamically orthogonal flow representation wit h uncertainty and a transformer model (encoder-decoder) for the path plann ing task. We model the problem as a sequence-to-sequence translation task where the source sequence is the agent's knowledge representation of the u ncertain environmental flow. The target sequence is the optimal sequence o f actions the agent must execute. We demonstrate that a trained transforme r model can predict near-optimal paths for unseen flow realisations and ob stacle configurations in a fraction of the time required by traditional pl anners. Validation is performed to show generalisation in unseen obstacle configurations. We also analyse the predictions of both transformer models \, viz\, decoder only and encoder-decoder and explain the inner mechanics of learning through a novel visualisation of self-attention of actions and states on the trajectories.\n\n\n\n\nALL ARE WELCOME CATEGORIES:Events,Thesis Defense END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20230726T100000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR